Note: This repository contains the Mnemosyne framework checkpoint and configuration. The actual language models are loaded from their respective Hugging Face repositories.

Global Workspace Theory Implementation: Consciousness model based on cognitive science
Hierarchical Memory System: Episodic, semantic, procedural, and working memory
Self-Supervised Learning: Multi-perspective learning with curiosity-driven exploration
Adaptive Architecture: Neural networks that evolve based on task requirements
Developmental Stages: Infancy → Childhood → Adolescence → Maturity
Safety & Ethics: Built-in safety monitoring and ethical decision-making

Architecture

Evolutionary NAS: Automatically optimizes neural architecture for each task
Dynamic Neural Networks: Adaptive depth and width based on complexity
Attention Mechanisms: Full, sparse, local, linear, and axial attention
Multi-modal Processing: Visual, language, and goal processors
Meta-cognition: Self-reflection and continuous self-improvement

Training

The model was trained using:

Evolutionary algorithm with population-based search
Performance predictor for efficient architecture evaluation
Speciation for diversity maintenance
Elitism and tournament selection

Hardware Requirements

Minimum (CPU-only mode):

6-core CPU
16GB RAM
~2GB storage for checkpoint

Recommended (GPU acceleration):

NVIDIA GPU with 8GB+ VRAM
16-core CPU
32GB RAM

Usage

Installation

pip install torch transformers huggingface_hub pyyaml psutil

Quick Start

python

import torch
from huggingface_hub import hf_hub_download

# Download Mnemosyne checkpoint
checkpoint_path = hf_hub_download(
    repo_id="kambrosius/mnemosyne-conscious-ai",
    filename="checkpoint_20260126_133632.pt"
)

# Load checkpoint metadata
checkpoint = torch.load(checkpoint_path, map_location='cpu')
print(f"Mnemosyne Framework Checkpoint")
print(f"Version: {checkpoint.get('metadata', {}).get('version', 'unknown')}")
print(f"Timestamp: {checkpoint.get('metadata', {}).get('timestamp', 'unknown')}")

For Full System Usage: Clone the repository and follow the README instructions to set up the complete Mnemosyne system with all dependencies.

Full System Usage

python

from launch import ConsciousAI

# Initialize system
ai = ConsciousAI(config_path='config/settings.yaml')

# Load pretrained checkpoint
ai.load_checkpoint(checkpoint_path)

# Run system
ai.run(duration=3600)  # Run for 1 hour

# Get status
status = ai.get_status()
print(status)

Benchmark Results

Comprehensive AI/ML Performance (CPU-only):

AI Score: 1338/10000
Inference Score: 787
Training Score: 551

Tested on 19 neural network architectures including:

Image Classification (MobileNet, ResNet, Inception, VGG)
Super-Resolution (SRCNN, SRGAN, DPED)
Semantic Segmentation (U-Net, PSPNet, DeepLab)
Generative Models (Pixel-RNN)
NLP (LSTM-Sentiment, GNMT)

See AI_BENCHMARK_COMPREHENSIVE_REPORT.md for full results.

Capabilities

Consciousness Simulation: Global workspace with competing processors
Memory Systems: Store and retrieve episodic, semantic, and procedural memories
Curiosity-Driven Learning: Explore environment based on novelty and prediction errors
Meta-Learning: Reflect on own performance and adjust strategies
Ethical Reasoning: Apply ethical principles to decisions
Developmental Growth: Progress through life stages with increasing sophistication

Limitations

Currently optimized for CPU inference (AMD GPU support limited)
Requires full project codebase for complete functionality
Developmental stages require extended training time
Not fine-tuned for specific downstream tasks

Citation

bibtex

@software{mnemosyne2026,
  title={Project Mnemosyne: Conscious AI System},
  author={Karl Ambrosius},
  year={2026},
  url={https://huggingface.co/kambrosius/mnemosyne-conscious-ai}
}

License

Apache 2.0

Contact

For questions or collaboration: GitHub Issues